What does konte do?
konte turns a short JSON spec into a storyboard image and a rough animatic, quickly enough to iterate on: the board takes a few seconds, and the video scales with its length and the machine. Nothing is generated by a model: a fixed renderer draws grey sets, one coloured mannequin per character, and the camera you described. The point is to see whether the idea works (who is where, what the camera does, how long each beat lasts) before spending money on real generation.
Run it as konte if that command exists. If it does not, use node "${CLAUDE_PLUGIN_ROOT}/cli.mjs" (keep the quotes: the path may contain spaces) in place of konte in every command below. It needs Node 22+, Chrome or Chromium, and ffmpeg for video. If Chrome is not found, set CHROME to its executable.
The loop
- Write the spec. Read
${CLAUDE_PLUGIN_ROOT}/docs/spec.mdonce, then write<name>.json. Start from the closest file in${CLAUDE_PLUGIN_ROOT}/examples/. Keep it small: 3 to 6 scenes, one sentence ofnoteper scene. - Lint it.
konte lint <name>.json. Fix everyerrorand everywarnyou did not intend. Lint runs on the same staging model the renderer draws, so trust it over your own estimate of the geometry; it samples and simplifies, so it does not replace reading the board. - Draw the board first.
konte render <name>.json --board-onlyis quick (a second or two for a few scenes) and writes only the board. Readboard-3d.png. Check each scene against its note: is the action visible, is the subject the right size, does the event read without the caption? - Revise the spec and repeat 2 and 3 until the board tells the story.
- Render the animatic.
konte render <name>.json. Show the person the board andsketch-3d.mp4, and say what you are unsure about. If they are going to a video model next, point them toprompts.mdand the clips inshots/. A full render removes clips and prompts left in that folder by an earlier render.
Do not skip step 3. A board is nearly free; a wrong video costs the person's attention.
Writing a spec that reads
- One idea per scene. If the note needs "and then", split the scene.
- Describe the shot, not the camera position. Use
size,angle,heightandmove. Reach for a raw camera only when those cannot say it. - Give each character one strong colour and keep it for the whole piece. Colour is how people are told apart across cuts.
- Make speeds real. Walking is about 1.4 m/s, running 3 to 8 m/s. A 16-second sprint covers about 100 m, so the set has to be that long.
- A jump lasts under a second. Use
"jump": [t0, t1, height]with theleapaction over the same time. For slow motion, give the jump its own short scene. - Close-ups need
part."size": "CU", "part": "hands"aims at the hands. Withoutparta close-up aims at the head. - A camera that stays put loses a moving subject. Use
trackorpanwhen someone travels. - Dialogue: seat people with
sit, add"look": "<other>", use"sit+talk"for the speaker, and cut between{"on": "a", "over": "b"}and{"on": "b", "over": "a", "side": "left"}. - Showing a product: make it a cast object with
"heldBy": "<person>", give the person"stand+show", and put the camera"on"the object for the close-up. - Give a screen, sign or logo board its own
color, or it disappears into a grey wall. - Set pieces are for geography, not decoration: the wall they run along, the gap they jump, the table between them. Three to eight pieces are usually enough.
- Use
beatswhen the moment that matters is not the one konte picked for the board. - Keep scenes short when the clips will feed a video model: 5 to 15 seconds, one scene per shot. Seedance 2.0 documents 15 seconds of reference video in total and Kling 4.0 3 to 30 seconds per clip.
What to tell the person
- The board and video are a sketch of staging and timing. Faces, lighting, texture and style are not in it and are still theirs to decide.
- The video has no words on it. Notes are on the board and in
prompts.md.--reviewwrites an extra copy with a slate, name tags and travel lines. shots/holds one clip per scene plus its first and last frame, andprompts.mdhas two prompts per shot. konte's clips and prompts have not been tested with any video model yet: say so, and present them as a starting point built from the models' documentation. With a model that takes a reference video (Seedance 2.5 or 2.0, Kling 4.0), attach the clip and use prompt A. With an image-to-video model that takes no reference video (Seedance 1.5 Pro), the model may keep the grey sketch look of its first frame, so suggest restyling the still into a finished image first and then using it as the first frame with prompt B. Ask which model they use before telling them which files to take; limits differ by model and platform.- Say plainly what the sketch cannot show (a facial expression, a product detail) rather than implying it is covered.